The United States is running an experiment no economy has run before: a single investment category, barely four years old, supplies half of measured GDP growth. Strip information-processing equipment and software — a sliver worth roughly 4% of GDP — out of the national accounts and the rest of the economy grew 0.1% annualized in the first half of 2025, on Jason Furman's pre-revision calculation. The question that matters for 2027 and beyond is not whether the buildout is large. It is what the buildout does to demand once the import bill is netted out — and what must replace it when the spending stops accelerating.
We decomposed the AI capex dollar into its four flows — foreign manufacturing, chip-design margin, domestic equipment and grid, construction — using customs data, company filings and the multipliers the literature supports. Three findings emerge. First, roughly 40 cents of each dollar leaves the country at the border, which is why the gross national accounts overstate the boom: the Federal Reserve Board's own July 2026 note warns that gross AI-investment contributions, unadjusted for net exports, overestimate the buildout's effect on GDP. Second, of what stays, nearly half lands as NVIDIA-type margin, which reaches demand almost exclusively through the stock market. Third, the arithmetic of the consensus capex path — ~$850 billion in 2026 rising to $1.4 trillion by 2029, with growth decelerating from ~100% to single digits — halves the investment channel's demand contribution by 2028. The earnings path the market is priced for requires AI adoption to fill that gap within 36 months, starting from a measured contribution of roughly zero.
Discussion